An Interpretable Alternative to Neural Representation Learning for Rating Prediction -- Transparent Latent Class Modeling of User Reviews

TL;DR

The paper tackles the need for interpretable rating prediction from textual reviews by introducing a Transparent Latent Class Model (TLCM) that organizes user and product latent classes on two 2D grids. Latent assignments are inferred with EM over a hierarchical latent structure that includes corrupted latent variables, enforcing a topographic prior via a neighborhood function $P( extbf{y}| extbf{z})$. An out-of-sample extension uses Bayes' rule to infer latent classes for unseen users from their reviews, and the inferred latent features feed a CNN-based rating predictor trained with Mean Squared Error. Empirical results on Amazon data demonstrate that TLCM yields competitive predictive performance while offering interpretable, visualization-friendly latent representations, addressing concerns about the interpretability gap in neural approaches. Overall, the work argues that principled, transparent models can closely approach, and in some settings match, the predictive power of text-based neural methods while providing clearer human-understandable explanations.

Abstract

Nowadays, neural network (NN) and deep learning (DL) techniques are widely adopted in many applications, including recommender systems. Given the sparse and stochastic nature of collaborative filtering (CF) data, recent works have critically analyzed the effective improvement of neural-based approaches compared to simpler and often transparent algorithms for recommendation. Previous results showed that NN and DL models can be outperformed by traditional algorithms in many tasks. Moreover, given the largely black-box nature of neural-based methods, interpretable results are not naturally obtained. Following on this debate, we first present a transparent probabilistic model that topologically organizes user and product latent classes based on the review information. In contrast to popular neural techniques for representation learning, we readily obtain a statistical, visualization-friendly tool that can be easily inspected to understand user and product characteristics from a textual-based perspective. Then, given the limitations of common embedding techniques, we investigate the possibility of using the estimated interpretable quantities as model input for a rating prediction task. To contribute to the recent debates, we evaluate our results in terms of both capacity for interpretability and predictive performances in comparison with popular text-based neural approaches. The results demonstrate that the proposed latent class representations can yield competitive predictive performances, compared to popular, but difficult-to-interpret approaches.

Figures (3)

• Figure 1: Input example for our architecture. Each pixel represents the value of the corresponding latent class probability.
• Figure 2: The schematic view of the proposed architecture. After the input layer, we follow the principles of related CNN-based models for rating prediction tasks.
• Figure 3: Results of the out-of-sample extension from the category Pet Supplies. The words in the bottom box are the ones used by the unknown user $u_{n}$ to review a product $p$ in our dataset. The considered words are included in the corresponding vocabulary $\mathcal{V}_{\text{pet}}$. Note that the darker the color, the higher is the probability assignment to the corresponding latent class.